Research Of 3D Reconstruction Based On Binocular Stereo Vision

Binocular stereo vision is one of the hotspots of computer vision. It perceives the world by simulating the human visual system. Binocular stereo vision is widely applied to many fields, such as games, ranging and 3D reconstruction. This paper mainly research the application of binocular vision in 3D reconstruction.The main work done in this paper are the following:(1) Analyzing the model of binocular stereo vision. Through the perspective projection model, the thesis analyze geometrical relationship of the three coordinate systems (image coordinate, camera coordinate and space coordinate).(2) Studying the principles of camera calibration. Camera calibration is the basis of binocular stereo vision. With the internal and external parameters obtained by calibration, the stereo matching can be greatly improved. According to the target of our research, this paper adopts a semi-automatic calibration proposed by Zhengyou Zhang which is different from the traditional calibration method and self-calibration. The thesis does the experiments of calibraion respectively based on Matlab and OpenCV. The experimental results show that Zhang’s calibration method is feasible and it can satisfy the accuracy of the thesis.(3) Focus on the analysis and research on the stereo matching algorithm. This paper analyzes the matching algorithm respectively from the similarity measure, the size of the support-window and the support-weight of neighborhood pixels. On above basis, this paper proposes the improvement strategy.On the aspect of the similarity measure, a improved Census transform is proposed which firstly transforms the image from RGB model to HS V model, and use the value of the model to replace the pixel gray value, the method can significantly reduce the effects of the external environment such as light.In addition, to determine the appropriate size of support window, a stereo matching algorithm of adaptive window is proposed which is based on the initial disparity map and edge detection in this paper. It can effectively solve the mismatching problem on the regions where the diaparities are discontinuous.The thesis proposed a novel adaptive support-weight stereo matching algorithm which obtains initial disparity map by traditional local algorithm, then by applying the similarity and proximity of the gestalt theory on the disparity map, it realizes the adaptive support-weight strategy of the aggregation window.The experiment result shows that the proposed algorithm has higher precision and efficiency than traditional local stereo matching algorithm.(4) Realizing the 3D reconstruction system based on binocular vision. The thesis Separately realize the camera calibration, stereo matching, acquiring the image correction and 3D spatial coordinates, and the visualization of three-dimensional scene reconstruction. We obtain the inner and external parameters of cameras by Zhang’s method. To gain disparity map, the proposed methods are adapted to the matching images corrected with camera parameters. The space coordinates of spatial points are acquired according to the triangulation principle. Finally, we adapt the Delaunay triangulation to the 3D scene space point cloud, then paste the texture and display the scene reconstructed by OpenGL.